I don't think this is really a good principal. Some optimizations work really well for some programs - should they not be implemented just because the compiler isn't one of those programs? Say the compiler gets no benefit from object-size improvements - does that mean no embedded software will get to benefit from those improvements?

Here's a principal I learned from a fairly awesome compiler writer, Mike Hind of Jikes RVM (paraphrasing):

"Students often ask me how much speed improvement an optimization should provide to be publishable. If it speeds programs up by 10%, but will take six months to implement, that's not a good result. But if it can be done in a day, and improves performance by 0.5%, that's a great result".